comparison notes/hnmf.m @ 19:f1f8c84339d0

Starting to revisit this EM logic and test in a single-column world

 
 % Iterate
 for it = 1:iter
     
     % E-step
-    zh = z .* permute(repmat(h,[1 1 R]),[3 1 2]);
+    zh = z .* permute(repmat(h,[1 1 R]),[3 1 2]); %% z is the source activation distribution, h the component (pitch) activation
     xa=eps;
     for r=1:R
         for k=1:K
             xa = xa + w(:,k,r) * squeeze(zh(r,k,:))';
         end;